Santander has teamed with NVIDIA to build new capabilities for evaluating artificial intelligence models and agents for banking, the bank announced on Monday (October 19). The work, led by the Santander AI lab, is designed to determine which AI model is best suited to a task and deploy it in the most efficient fashion.
The centrepiece is a testing setup Santander calls its “Model Arena.” Rather than defaulting to the largest general-purpose model available, the bank says it has been progressively moving towards smaller, specialised models designed for specific banking tasks, and it is using NVIDIA software to evaluate and tune them.
Model Arena: How Santander Matches A Model To Each Banking Task
Santander said it has constructed its Model Arena to evaluate open-weight models against banking tasks and determine which model is right for each workload. The approach treats model selection as an engineering decision: a model is scored on the bank’s own tasks before it is put into production.
That matters because a global bank runs very different kinds of AI work side by side, from customer conversations to document handling and internal decisioning. A single large model serving all of it is simple to buy, but Santander’s lab is betting that a bench of smaller models, each specialised for one job, will be cheaper to run at bank scale.
The bank is specialising those models using NVIDIA NeMo Customizer, part of NVIDIA’s NeMo software family for customising models with an organisation’s own data and workflows. It is the same enterprise software layer that sits behind why NVIDIA powers the AI boom: the chips draw the headlines, but the tooling is what large customers use day to day.
The Cost Numbers: From €5 Per Interaction To About €0.03
In selected experiments, Santander said its specialised-model approach cut serving costs from €5 per interaction to approximately €0.03, while upholding strong performance. The bank did not present the figure as an average across its operations; it is a result from selected experiments, reported in its own release.
The partnership is already producing production-style results, the companies say. In one experimental customer-service use case, Santander improved conversation quality while lowering costs by more than 50 percent, by deploying a more efficient AI model.
“In this new world, cost per intelligence output is a key metric,” said Aser Blanco, NVIDIA’s global banking lead. “Santander is showing what that looks like in production with smaller models, specialized on the bank’s own data, evaluated against real banking tasks and served efficiently enough to run at the scale of a global bank.” Blanco noted that banking is one of the most data-centric industries, and that the value lies in turning that data into proprietary intelligence through training and fine-tuning rather than model size alone.
What Santander And NVIDIA Say Comes Next
Santander and NVIDIA say they will continue to work together on foundation models, AI agents, fraud and anti-money-laundering capabilities, real-time decisioning and advanced simulation technologies, such as banking digital twins.
The agenda lands in a sector that has already moved fast. Recent research by PYMNTS Intelligence found that 95 percent of large financial services firms surveyed had widely deployed or embedded newer AI tools in their data and technology processes, while 30 percent of financial services companies pointed to data quality and fragmentation as their main obstacle to further deployment.
It also lands a day before NVIDIA’s GTC Berlin conference opens, where the company’s enterprise and sovereign-AI push will be on stage, and in the same week that Samsung and TSMC were reported to be working on NVIDIA’s next-generation memory. Santander’s announcement is the software-and-services side of that same expansion.
Why It Matters
The central entity here is the Santander–NVIDIA banking deployment, and its defining attribute is cost per interaction. A bank that can evaluate open-weight models against its own tasks, specialise them with NeMo Customizer and serve the result for cents instead of euros has a different AI budget conversation from a bank renting a frontier model by the token for every job.
For NVIDIA, the deal is another proof point that its AI business is not only silicon: how NVIDIA makes money from AI increasingly runs through software platforms that keep customers inside its stack. For rival banks, the Model Arena is a template worth copying or beating. And for the open-weight model ecosystem, a top-tier European bank publicly benchmarking downloadable models against real banking work, and choosing them, is exactly the kind of enterprise validation those models have been chasing.
Background
Santander, headquartered in Spain, is one of Europe’s largest banks, with retail and commercial operations across Europe and the Americas. NVIDIA spent the past decade growing from a graphics-chip maker into the central supplier of AI computing, and its NeMo family is the software layer it offers enterprises that want to adapt models rather than just rent them. Banks, which sit on large volumes of proprietary customer and transaction data, have become a prime battleground for that pitch, because their data cannot simply be handed to a general-purpose model without careful evaluation first.
Source: PYMNTS; Finimize (Reuters)
0 Comments